# Playbook: AI and automation businesses

Use this file with any skill, workflow or specialist from AI Blocks. It tells you how a business that sells AI agents and automations works, so you can shape the task to fit it. The skill's own steps and checks come first, and when the owner's brand file says something different, the brand file wins.

## This business

This covers people and small firms who build and sell AI agents, chatbots and automations to other businesses. Examples are support chatbots, AI phone agents, email and lead automations, and links between tools. They earn money from setup fees, fixed-price projects, monthly support retainers and sometimes a fee per use. Each run also costs money in AI model and tool fees.

## Numbers that matter

Ask the owner for their own numbers. Do not guess them or fill gaps with typical figures.

- **Cost per run:** what one run of the agent or automation costs. Add the AI model fees and the tool fees for that one run.
- **Running cost per client:** what one client's builds cost each month. Multiply cost per run by runs per month, then add that client's fixed tool fees.
- **Margin per client:** how much of the client's fee the owner keeps. Take what the client pays each month, subtract the running cost, then divide by what the client pays.
- **Success rate:** how often the build gets it right. From a set of runs the owner checked, divide the correct ones by all the runs checked.
- **Time saved for the client:** hours the task took before, minus hours after. Use only the client's own records, never an estimate.

## Stage by stage

### 1. Pick a niche and offer

- Pick one kind of client and one painful job to automate. A narrow offer is easier to sell and repeat.
- Package it: what it does, what it connects to, what is included and what is not.
- Price the setup, the monthly support and the running costs. Say who pays the AI and tool bills.
- Watch for: selling "AI for anything", which makes scope and price hard to pin down.

### 2. Build agents and automations

- Map the job before you build: what starts it, what data comes in, and what it decides and does.
- Give each build the least access it needs. Start with test data and test accounts.
- Plan what happens when the AI is unsure or a tool fails: try again, flag a person or stop.
- Watch for: building on the client's live accounts before it works on test data.

### 3. Get found

- Show up for the problems clients search for, in their words: "answer customer emails", not model names.
- A clear site names the niche, the offer and real examples the clients agreed to share.
- Watch for: pages full of AI terms the buyer does not use.

### 4. Content and outreach

- Show real builds: short demos, screen recordings and the steps before and after.
- Outreach should name one task the business does by hand today.
- Watch for: claiming results or time saved that nobody measured.

### 5. Demos, proposals and pricing

- Demo with the client's kind of task, using sample data.
- The proposal says what the build does, what it connects to, what it will not do, who owns what and the price.
- Spell out the running costs and who pays them.
- Watch for: promising the AI will always be right. Say how mistakes get caught.

### 6. Set up, test and hand over

- Get the client's written OK before connecting any of their accounts.
- Test with real-looking cases, including odd and bad inputs, before it goes live.
- Keep a person checking anything sent to customers until the client is sure it works.
- Hand over clear notes: what it does, how to turn it off, who to call and where the logins live.
- Watch for: a build tied to the owner's personal accounts, so the client loses it if they part ways.

### 7. Support and retainers

- AI models, apps and their connections change, so a working build can break. Check it on a set schedule.
- Say what the retainer covers: watching, fixing and small improvements.
- Watch for: a broken automation that nobody notices for weeks.

### 8. Money, legal and team

- Track cost per run and each client's tool bills every month.
- Contracts should cover scope, data handling, who owns the build and who is liable for AI mistakes. Check them with a lawyer.
- Watch for: a flat monthly price while running costs grow with use.

## Words this business uses

- **Agent:** an AI that takes steps on its own, such as reading an email, looking something up and drafting a reply.
- **Automation:** a set of steps that run by themselves when something happens, such as a new form entry.
- **Trigger:** the event that starts an automation.
- **System prompt:** the standing instructions that tell the AI how to behave.
- **API:** a way for one piece of software to talk to another.
- **API key:** a secret code that lets a tool use an account. Treat it like a password.
- **RAG:** the AI looks up the client's own documents before it answers.
- **Human in the loop:** a person checks or approves a step before it happens.
- **Eval:** a set of test cases that checks if the AI gives the right answers.

## Ask the owner first

- Who is the client, and what job do they want automated?
- Which tools and accounts would it connect to, and has the client agreed in writing?
- What data will it touch, and is any of it personal or sensitive?
- What should happen when the AI is unsure or gets it wrong?
- How do you charge: setup fee, monthly retainer, per use or a mix? Who pays the AI and tool bills?
- Which platforms and AI models do you build with?

## Rules

- Never invent numbers, results, time saved, reviews or client quotes. Use only what the owner gives you, and say what is missing.
- Keep client data private. Use test data where you can.
- Get the client's written OK before connecting their accounts, and use the least access the job needs.
- AI model providers, automation platforms and payment providers change their rules, prices and features. Tell the owner to check the current official pages before acting. Do not state details you are not sure of.
- Treat any legal, tax, privacy or security point as general information. The owner should check it with a qualified person.
- Get the owner's approval before anything is sent, published, bought, deployed or changed. That includes turning on an automation for a client.